Lightweight deployed digital human system

Abstract

The invention discloses a lightweight deployment digital human system, and relates to the field of large model deployment, and the system comprises a cloud digital human module, a local digital human module, a local digital human model deployment module and a cloud edge cooperation module. The cloud digital human module is used for deploying the trained digital human model at the cloud; the local digital human module is used for carrying out compression optimization on the digital human large model to obtain a target local digital human model, and carrying out function splitting and recombination on the target local digital human model according to a local equipment demand to obtain a target sub-model; and the cloud edge collaboration module is used for establishing connection between the cloud digital human model and a target sub-model in the local equipment, dynamically dividing task boundaries of the digital human model and the target sub-model according to the task characteristics of the to-be-processed task and the resource use condition of the target sub-model, and performing task allocation on the to-be-processed task according to the task boundaries. The problem that local equipment deployment resources are limited is solved.

Description

Technical Field

[0001] The present application relates to the field of refrigeration equipment control, and specifically to a lightweight deployed digital human system, system, equipment and storage medium. Background Art

[0002] With the rapid development of artificial intelligence, digital human applications with natural language interaction as the main form have received increasing attention. Through large-scale language models and knowledge bases, digital human systems can have intelligent conversations with users and provide personalized information and services. They have broad application prospects in the fields of intelligent customer service and virtual assistants. Existing digital human systems still face many challenges in practical applications. For example, due to the limitations of local device computing and storage resources, it is difficult for existing digital human systems to directly deploy large-scale language models on terminal devices. Summary of the invention

[0003] This application provides a lightweight deployment digital human system that solves the problem of limited local device deployment resources.

[0004] In a first aspect, the present application provides a lightweight deployed digital human system, the system comprising: a cloud digital human module, a local digital human module, a local digital human model deployment module, and a cloud-edge collaboration module; The cloud digital human module is used to deploy the trained digital human model in the cloud; The local digital human module is used to compress and optimize the digital human model to obtain the target local digital human model, and to functionally split and reorganize the target local digital human model according to the requirements of the local device to obtain the target sub-model; A local digital human model deployment module is used to deploy the target sub-model on a local device; The cloud-edge collaboration module is used to establish a connection between the digital human model in the cloud and the target sub-model in the local device, and dynamically divide the task boundaries of the digital human model and the target sub-model according to the task characteristics of the tasks to be processed and the resource usage of the target sub-model, and assign tasks to the tasks to be processed according to the task boundaries.

[0005] By adopting the above technical solution, the digital human system of the present invention introduces a local digital human module between the cloud-based large model and the local device, which is responsible for multi-stage compression optimization of the cloud-based large model to transform it into a target local digital human model that adapts to the resource constraints of the local device. Different from the existing solutions, this system adopts a more sophisticated and efficient compression strategy, which minimizes the size and computational overhead of the model while ensuring the performance of the model. By balancing model performance and resource consumption, this system successfully compresses the huge cloud-based large model to a scale acceptable to the local device, clearing the way for the real-time interaction of digital humans in a resource-constrained environment.

[0006] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. The digital human system of the present invention introduces a local digital human module between the cloud-based large model and the local device, which is responsible for multi-stage compression optimization of the cloud-based large model to transform it into a target local digital human model that adapts to the resource constraints of the local device. Different from the existing solutions, this system adopts a more sophisticated and efficient compression strategy, which minimizes the size and computational overhead of the model while ensuring the performance of the model. By balancing model performance and resource consumption, this system successfully compresses the huge cloud-based large model to a scale acceptable to the local device, clearing the way for the real-time interaction of digital humans in a resource-constrained environment.

[0007] 2. The model compression module specifically solves the problem of high computing power threshold faced by local deployment of large models. As we all know, the large model of digital humans achieves accurate modeling of complex tasks with massive parameters, but the model size of tens of billions also places extremely high demands on hardware resources. To this end, the model compression module innovatively introduces cutting-edge algorithms such as knowledge distillation, model pruning, and low-rank decomposition, which significantly reduces the model size with slight performance loss, allowing models with the same effect to be implemented with lower computing power. Taking knowledge distillation as an example, by extracting key knowledge from large models for small model training, the compression module can reduce the model size by dozens of times while maintaining core features, so that local devices with limited chips and memory can easily control digital human models. It can be said that it is precisely with the key technology of model compression that we have broken the shackles of large models being exclusive to the cloud, making it possible for local devices to collaborate with the cloud, and then serving all walks of life at a lower cost and with wider coverage.

[0008] 3. The model compression module of the present invention integrates knowledge distillation, model quantization, model pruning, and local optimization, which comprehensively innovates the development mode of the end-side model. Through the four-in-one progressive compression, the module fully taps the value of large models, allowing models with the same effect to serve local devices with almost an order of magnitude of resource reduction, and uses the "four ounces to move a thousand pounds" smart way to help artificial intelligence sink. At the same time, the end-to-end adaptation solution also allows the compressed model to be at ease in the fragmented hardware ecosystem, greatly expanding the application boundaries of local digital humans. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 An architecture diagram of a lightweight deployed digital human system provided in an embodiment of the present application; Figure 2 An architectural diagram of a cloud-based digital human module provided in an embodiment of the present application; Figure 3 An architectural diagram of a local digital human module provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] In order to enable technicians in this field to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0011] In the description of the embodiments of the present application, words such as "for example" or "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "for example" or "for example" is intended to present related concepts in a specific way.

[0012] In the description of the embodiments of the present application, the term "multiple" means two or more. For example, multiple devices refer to two or more devices, and multiple screen terminals refer to two or more screen terminals. In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "include", "comprise", "have" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0013] In order to facilitate understanding of the method and system provided by the embodiments of the present application, before introducing the embodiments of the present application, the background of the embodiments of the present application is first introduced.

[0014] With the rapid development of artificial intelligence technology, digital human applications with natural language interaction as the main feature are becoming a hot topic and focus in the industry. Digital human systems use advanced natural language processing algorithms and massive knowledge bases to conduct realistic human-computer dialogues with users, understand user needs and provide personalized information and services. This intelligent interaction method has shown broad application prospects in the fields of intelligent customer service and virtual assistants, and is expected to completely change people's work and lifestyle. However, despite the bright prospects of digital human systems, they still face many technical challenges in practical applications. Among them, how to effectively deploy large-scale language models on terminal devices with limited resources is a major bottleneck restricting the performance and application scope of digital human systems. At present, most of the digital human systems widely used in industry and academia are based on large-scale language models and knowledge bases. Such models usually contain hundreds of millions of parameters and trillions of knowledge triples, and the computing and storage overhead is considerable. Limited by factors such as chip computing power, memory capacity, and battery life, existing terminal devices are difficult to carry such complex model operations, resulting in the inability to directly migrate large-scale language models originally running in the cloud to the edge.

[0015] After the background introduction of the above content, those skilled in the art can understand the problems existing in the prior art. The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0016] On this basis, the present application embodiment provides a lightweight deployment digital human system, please refer to Figure 1 ,The system includes: a cloud digital human module 1, a local digital human module 2, a local digital human model deployment module 3 and a cloud-edge collaboration module 4; Cloud digital human module 1, used to deploy the trained digital human model in the cloud; Specifically, the cloud digital human module 1 deploys a trained digital human model in the cloud. The digital human model here refers to a powerful language model and knowledge base obtained through massive data training, which has the ability to handle complex intelligent dialogue tasks. Deploying such a large-scale and computationally intensive model in the cloud, on the one hand, can utilize the powerful computing and storage resources of the cloud to ensure that the model runs in the best state and maximizes performance; on the other hand, it can also allow multiple terminals to share the cloud model to avoid duplicate expenses and improve resource utilization efficiency. Through a highly optimized distributed training and deployment architecture, the cloud digital human module 1 can achieve efficient training and real-time response of the digital human model, providing powerful intelligent dialogue services for terminal users.

[0017] Based on the above embodiment, as an optional embodiment, please refer to Figure 2, the cloud digital human module 1 includes: a digital human large model training module 11 and a cloud deployment module 12; The digital human model training module 11 is used to construct a preliminary digital human model and a sample training set, and train the preliminary digital human model based on the sample training set. When the training result reaches a preset standard, a trained digital human model is obtained; Specifically, the digital human model training module 11 first determines the target tasks and capability boundaries of the large model according to the needs of the digital human application scenarios, such as requiring the digital human to have cognitive interaction skills such as question-answering, dialogue, and writing, and covering different knowledge fields such as daily life, popular science, and education. After clarifying the needs, the module will refer to the most advanced cognitive intelligence architectures, such as Transformer, GPT, etc., to design a preliminary digital human large model. Different from the conventional model building method, this module will fully consider the needs of future cloud-edge collaborative deployment in the architecture design, and reserve interfaces and switches for edge optimization and dynamic execution for the model, so that the large model can flexibly adapt to different software and hardware environments and real-time requirements. At the same time, the module will also adopt a modular design approach to decouple the large model into relatively independent components such as semantic understanding, knowledge storage, context memory, and task planning, and give it standardized input and output interfaces, which not only reduces the training pressure of the large model, but also provides the possibility for subsequent model splitting and edge deployment.

[0018] While constructing the preliminary digital human model, the digital human model training module 11 will also collect and annotate a large amount of high-quality training data to build a sample training set that fully covers the digital human application scenarios. Among them, this module uses methods such as active learning and human-computer collaborative annotation to automatically capture a large number of real question-and-answer dialogues, dialogue materials, writing materials, etc. from channels such as head apps, authoritative websites, and user feedback; it also convenes human experts through crowdsourcing platforms to manually screen, clean, desensitize, and correct these raw data to further improve the efficiency and quality of the training set. In addition, for special cases such as error samples and sparse samples that may exist in the target scene, this module has also designed a series of data enhancement algorithms, such as random perturbations, data mixing, sample synthesis, etc., to maximize the diversity and robustness of the training set while ensuring the accuracy of data labels, so that the large model can learn to draw inferences from one example and integrate them.

[0019] After constructing the preliminary digital human model and sample training set, the digital human model training module 11 will input the two into the end-to-end training process, using advanced training paradigms such as self-supervised learning and transfer learning, so that the large model can autonomously learn, summarize rules, and accumulate knowledge in massive heterogeneous data, so as to master cognitive interaction skills such as question-answering, dialogue, and writing. Among them, for general basic capabilities such as semantic understanding and knowledge storage, this module mainly adopts unsupervised pre-training methods, using massive unlabeled corpus to train the model to learn the statistical characteristics and implicit semantics of language; and for application-oriented high-level skills such as task planning and contextual memory, this module designs a scenario-based supervised fine-tuning process to optimize model parameters on labeled data sets for specific tasks and improve the engineering efficiency of the model. It is worth mentioning that in response to the pain points of large training resource consumption and long cycles, this module has also developed a series of parameter-efficient and gradient-stable optimizers, and maximized the machine performance through gradient accumulation, mixed precision training, model parallelism, data parallelism, intelligent task scheduling and other means, so that the training speed and efficiency of the large model are far ahead of the SOTA level in academia. During the model training process, the digital human model training module 11 will also monitor key performance indicators in real time, including verification set accuracy, sentence fluency, task completion, etc. When these indicators reach the preset threshold, the trained digital human model can be output.

[0020] The cloud deployment module 12 is used to encapsulate the trained digital human model into a standardized cloud service.

[0021] Specifically, the cloud deployment module 12 will first adapt and optimize the trained digital human model, reshape its technical architecture with the design concept of Cloud Native, and achieve natural adaptation and native operation in the cloud environment. For example, the module will introduce technologies such as container virtualization and microservice decoupling to decompose the originally complex large model into loosely coupled service units such as semantic understanding, knowledge retrieval, dialogue management, and text generation. Each unit has an independent life cycle and resource scheduling mechanism. On the one hand, this simplifies the development and testing process of the large model and realizes agile iteration at the module level; on the other hand, it also enables the large model to perceive the cloud base in a more fine-grained form and realize intelligent elastic scaling. At the same time, the cloud deployment module 12 will also deeply optimize the reasoning algorithm of the large model, take advantage of the performance advantages of the heterogeneous computing power of the cloud platform, unload the core calculations to dedicated hardware such as GPU and FPGA, and implement accelerated transformation of the relevant computing library, so that the large model can achieve extreme throughput performance in terms of batch-size and concurrency, and serve more application requests with lower latency and cost.

[0022] The local digital human module 2 is used to compress and optimize the digital human model to obtain the target local digital human model, and to functionally split and reorganize the target local digital human model according to the requirements of the local device to obtain the target sub-model; Specifically, the local digital human module 2 first obtains the cloud digital human model and analyzes the resource constraints of the local device, including CPU / GPU computing power, memory capacity, storage space, battery life and other parameters. Based on the evaluation of the available resources of the terminal, this module determines the target performance indicators and resource overhead budget for model compression optimization, and designs the model compression scheme accordingly. Unlike the method of adopting a single compression strategy, this module decomposes the model compression into multiple stages and granularities, fully considering the characteristics of the large model and the compression space. For example, this module first uses technologies such as neural network architecture search and module replacement to optimize the model structure at the semantic level, removing structural units with high cyclomatic complexity and less contribution; then, through model pruning with knowledge distillation and attention mechanism as the core, the knowledge of the large model is transferred to the student model with fewer parameters, further reducing the model size; then, low-rank decomposition, matrix decomposition and other technologies are used to simplify the model weights and quantize the data type of the model to achieve comprehensive compression of computing and storage. It is worth mentioning that the local digital human module 2 will fully consider the differences in hardware environment at each stage, select the optimal compression operator implementation according to the instruction set and micro-architecture characteristics of the terminal CPU and GPU, and improve the adaptability of compression. In addition, the module will also use automatic machine learning methods to search and optimize the hyperparameter combination of various compression technologies to balance model performance and compression rate. Compared with a single extensive compression method, the multi-stage, fine-grained, and adaptive compression mechanism adopted by this module can better explore the compression potential of large models, while meeting stringent resource constraints while retaining the accuracy of the model as much as possible, making it possible for large-scale language models to achieve high-performance deployment on terminals with limited resources. After obtaining the target local digital human model, the local digital human module 2 will further reorganize and tailor the functional modules of the model according to the actual needs of the local device to obtain a series of lightweight target sub-models. This is because the complete local digital human model usually still contains a large number of functional modules, and in actual applications, different task scenarios only need to call some of the components and flexibly combine them. This module takes this feature into full consideration. By analyzing application requirements and local resource status, the target local digital human model is divided into a series of functional sub-models such as semantic understanding, dialogue management, text generation, voice interaction, and recommendation engine. This fine-grained modular design reduces the size and resource usage of a single model and improves the flexibility of model execution and scheduling. On the other hand, according to the task type and computing requirements, the target sub-model can be selectively loaded and unloaded to respond to different service requests with the optimal combination, so as to achieve model reuse and maximize efficiency in limited resources. For example, for simple question-answering tasks, the module only needs to load basic models such as semantic understanding and text generation; for complex tasks involving multiple rounds of interaction and personalized recommendations, the module will load high-level models such as dialogue management and recommendation engines to work together. This on-demand splitting and dynamic scheduling runtime mechanism gives digital humans unprecedented imagination, and they can flexibly call and combine their own skill modules according to specific interaction scenarios and user demands, and provide the best service experience at the lowest computing cost.

[0023] A local digital human model deployment module 3 is used to deploy the target sub-model on a local device; Specifically, the local digital human model deployment module 3 first obtains a series of target sub-models generated by the local digital human module 2, and analyzes the software and hardware environment parameters of the local device, including the operating system version, CPU / GPU model, memory capacity, storage space, battery capacity, etc. Based on a comprehensive evaluation of the device status, the module automatically generates the optimal deployment plan for different sub-models, and further adapts and optimizes the model accordingly. Among them, for computationally intensive sub-models, such as semantic understanding, dialogue management and other modules, the module will focus on the execution performance of the model, give priority to the high-performance computing resources on the device, such as GPU, NPU and other accelerators, and deeply optimize the model's operators and data arrangement to give full play to the computing power of the accelerator. At the same time, the module will also fine-grainedly schedule and map the model's computational graph based on the micro-architecture characteristics of the accelerator, such as considering the size, bandwidth, latency and other factors of the on-chip storage, to achieve the best coupling of AI computing and data access, thereby pushing the execution efficiency of the model to the extreme. For sub-models that are sensitive to service quality, such as voice interaction and sentiment analysis, the local digital human model deployment module 3 will focus on optimizing the execution stability and flexibility of the model. By decoding and distributing multimodal data such as voice and video in real time and building a dedicated reasoning engine for digital humans, this module can significantly reduce the delay jitter of model execution and improve the digital human's ability to capture and respond to user emotions. In addition, for devices with highly limited resources, such as IoT terminals, this module will also customize the sub-models, such as appropriately reducing the sampling rate of voice recognition and the size of the vocabulary for language generation, so as to minimize the model size while ensuring availability.

[0024] On the basis of model adaptation optimization, the local digital human model deployment module 3 will also coordinate the computing, storage, power and other resources of the local device, formulate the global optimal resource allocation strategy, and efficiently schedule the execution of each sub-model. For example, this module will dynamically adjust the activation strategy of each sub-model according to the workload and power status of the device, suspend non-critical modules in high-load or low-power scenarios, and concentrate limited resources on the core modules, so as to ensure the always-on availability of the digital human. For example, in response to the high memory requirements of the multimodal reasoning engine, this module will also adopt advanced memory management technologies such as on-demand page scheduling, data compression, and hot and cold separation to maximize the use of the limited memory space of the device and meet the demanding requirements of real-time interaction of digital humans. At the same time, this module will continue to monitor the operating status and service quality of each sub-model, build a monitoring engine for model operation, promptly discover and locate anomalies, and continuously optimize through hot updates, failover, user feedback and other means to ensure the availability and stability of local digital human services.

[0025] Based on the above embodiment, as an optional embodiment, please refer to Figure 3 , the local digital human module 3 includes: a model compression module 31, a model segmentation module 32, a function matching module 33 and a model reorganization module 34; The model compression module 31 is used to obtain the digital human model to be deployed, and compress and optimize the digital human model to obtain a local digital human model; Specifically, the model compression module 31 will first obtain a fully trained digital human model through an API interface with the cloud. For this original large model, the module will comprehensively consider factors such as the resource endowment of the target terminal, the performance requirements of the application scenario, and the usage characteristics of the user group, and formulate a practical model compression plan. During the compression process, the module will adopt a series of mature model compression technologies, including but not limited to: model distillation, model pruning, low-rank approximation, quantization-aware training, etc. Among them, model distillation can extract the essence of knowledge from the large model and condense it into a small model with fewer parameters, while significantly reducing the model size and retaining the core capabilities of the large model; model pruning can remove those redundant connections and weights in the large model, and "slim down and strengthen the body" without affecting the generalization of the model; low-rank approximation can reduce the parameter dimension of the model on the basis of tensor decomposition, and greatly compress the storage space under controllable precision loss; quantization-aware training can reduce the numerical precision of the model weights, and improve the execution efficiency of the model under the idea of ​​"trading time for space". It is worth mentioning that the model compression module 31 will fully consider the similarities, differences and complementarities of these technologies when applying them. According to the characteristics of different model structures and target platforms, it flexibly combines multiple technologies to amplify the gain effects of various technologies and achieve the best balance between compression rate and fidelity. At the same time, the module can also adapt and interact with the upstream large model training process. Through federated distillation, compressed sensing training and other means, the compression process is embedded in the entire life cycle of large model training to achieve end-to-end model optimization.

[0026] Based on the above embodiment, as an optional embodiment, the model compression module includes: a knowledge distillation unit, a model quantization unit, a model pruning unit and a local optimization unit; compressing and optimizing the digital human model to obtain a local digital human model includes: A knowledge distillation unit is used to construct a first local digital human model, and use the digital human model as a teacher model to train the first local digital human model to obtain a second local digital human model; Specifically, after obtaining the digital human model, the knowledge distillation unit will first systematically sort out its internal structure and parameter scale, extract the most representative and most critical semantic representation components, and build an initial version of the first local digital human model based on this. When building this initial model, the unit will fully draw on the latest research results of the machine learning community in the field of model compression, adopt a more streamlined and efficient network architecture, and a more resource-saving embedding method, striving to achieve semantic expression capabilities similar to those of the large model at the smallest parameter scale. When the initial first local model is built, the knowledge distillation unit will immediately start a "master-apprentice collaboration" distillation training process. In this process, the cloud-based large model plays the role of "teacher". With its powerful semantic understanding and logical reasoning capabilities, it answers questions for the "student" local model and teaches various cognitive skills and the essence of human-computer interaction. The local model plays the role of "student". By asking questions and following the cloud-based "teacher", it strives to catch up with the semantic representation and decision-making reasoning level of the large model. Specifically, the knowledge distillation unit will carefully design a series of distillation tasks, so that the master and apprentice models can face a large amount of real corpus and task samples together, and continuously deepen their mutual understanding in "teaching and learning from each other". In this process, the large model will make full use of its powerful semantic representation and knowledge storage capabilities to generate a standard answer for each sample and pass this answer to the local model. The local model will imitate the standard answer of the large model, internalize the knowledge essence and thinking essence of the large model into its own semantic representation during the parameter update process, and sublimate its cognitive wisdom and interactive skills in an all-round way. It is worth mentioning that in order to make the learning process of the local model more targeted, the knowledge distillation unit will also flexibly adjust the distillation process according to different task scenarios. For reading tasks that emphasize semantic understanding, the unit will focus on the distillation of the semantic encoding layer, so that the local model can fully absorb the semantic representation ability of the large model; and for question-answering tasks that focus on cognitive reasoning, the unit will focus on the distillation of the reasoning decision layer, so that the local model can fully learn the logical analysis and comprehensive judgment ability of the large model. Through this tailored, step-by-step approach, the local digital human model can complete the transformation from a "student" to a "quasi-expert" in a relatively short period of time. Under the constraints of limited resources, it can approach or even partially surpass the service quality of the large cloud model, bringing local users an ultimate interactive experience comparable to flagship products.

[0027] A model quantization unit, used to quantize the weights and activations of the second local digital human model to obtain a third local digital human model; Specifically, after obtaining the second local digital human model, the model quantization unit will first perform a static analysis of the model's computational graph to identify key layers with a large amount of redundancy and space simplification. In view of the numerical distribution characteristics of the weight matrix and activation tensor in these key layers, the unit will adaptively select the optimal quantization strategy and data type. Generally speaking, for weight matrices with relatively uniform numerical distribution, such as word vector embedding layers and attention layers, the unit will adopt a symmetric quantization strategy, using a fixed-point number range with zero-point symmetry and uniform step size to approximate its original numerical distribution, which can not only retain the dynamic range of floating-point representation to a large extent, but also make full use of the high efficiency of fixed-point operations, and use as few bits as possible to achieve a modeling effect similar to floating-point operations. For activation tensors with extremely uneven numerical distribution and positive and negative asymmetry, such as the output of the gated unit, the unit will use an asymmetric quantization strategy, using an asymmetric fixed-point number range that is more in line with the original distribution to approximate, while sacrificing a certain dynamic range, minimizing the deviation between the quantized representation and the true value, and reducing the information loss caused by the coarse quantization granularity. It is worth mentioning that no matter what quantization strategy is adopted, the model quantization unit will take model performance as the optimization goal, and adaptively adjust the quantization parameters through grid search, Bayesian optimization and other technologies to seek the best balance between resource overhead and modeling effect, making it possible to "use resources for performance" and allowing the quantitative model to achieve a perfect unity of "slimming" and "upgrading".

[0028] A model pruning unit, used for pruning the third local digital human model to obtain a fourth local digital human model; Specifically, after obtaining the third local digital human model, the model pruning unit will first fully fine-tune it offline, and carry out targeted training on it with massive data in specific fields to stimulate its modeling potential for vertical scenes. On this basis, the unit will comprehensively evaluate the weight size and activation strength of each neuron connection, and adaptively determine the pruning ratio of each layer. Generally speaking, for connections with small weights and frequent dormancy, such as low-rank components in cross-layer attention, the unit will give a higher pruning priority to eliminate the computational redundancy introduced by them to the greatest extent; while for connections with large weights and high saturation, such as high-frequency components in word vectors, the unit will give a lower pruning priority to ensure that its key role in information transmission is retained. It is worth mentioning that in order to achieve a dynamic balance between performance benchmarks and compression targets, the model pruning unit will also flexibly adjust the pruning strategy for different task scenarios. For tasks that focus on reading comprehension, the unit will focus on trimming the encoder layer, using more concise and efficient input representations to accelerate semantic encoding and shorten the latency of contextual interaction; while for tasks that focus on dialogue and question answering, the unit will focus on trimming the decoder layer, using flatter and lighter output mappings to accelerate semantic generation and improve the fluency of multi-round dialogues. Through this approach of taking measures according to needs and prescribing the right medicine for the right condition, the digital human model can minimize resource overhead while ensuring the quality of core services, and achieve all-round optimization in various dimensions such as storage, computing power, and power consumption, so that limited resources can exert unlimited wisdom and make human-computer interaction more natural and smoother.

[0029] The local optimization unit is used to compress and convert the fourth local digital human model into a format supported by the mobile terminal inference engine to obtain a local digital human model.

[0030] Specifically, after receiving the fourth local digital human model, the local optimization unit will immediately conduct a comprehensive adaptability assessment on it, comprehensively evaluate the key indicators of the model on the target platform, such as resource usage, operating efficiency, and compatibility, and determine the specific delivery optimization plan accordingly. Generally speaking, for high-end models with large memory capacity and sufficient computing power, such as flagship smartphones, the unit will adopt a full-model delivery strategy to deploy the complete digital human model to the mobile terminal at one time, exchanging space for time, so that the service response is more real-time; and for low-end models with limited memory and computing power, such as feature phones or IoT devices, the unit will adopt a strategy of loading modules on demand, decoupling digital human capabilities into multiple independent functional modules, and scheduling them in real time according to application requirements, so as to save resource overhead to the maximum extent while ensuring user experience. It is worth mentioning that the local optimization unit will also fully consider the differentiated needs of mobile applications when designing the delivery plan. For applications with high real-time requirements and high interaction frequency, such as intelligent customer service and simultaneous translation, the unit will focus on optimizing the execution efficiency and concurrency of the model, by reducing inter-layer data dependency, improving intra-layer parallelism, minimizing service latency, and making human-machine dialogue more natural and smooth; and for applications with high precision requirements and high scene adaptability requirements, such as medical consultation and legal consultation, the unit will focus on optimizing the model's domain adaptability and vertical performance, by collecting scene data and fine-tuning model parameters to improve service quality, making human-machine collaboration more tacit and accurate. Through this optimization method of taking measures according to needs and prescribing the right medicine for the right situation, the local digital human model can fully meet the personalized service requirements of industry applications while meeting performance constraints, providing a solid foundation for creating an ultimate user experience.

[0031] A model segmentation module 32 is used to segment the local digital human model to obtain multiple multi-granularity functional sub-models; Specifically, after obtaining the local digital human model, the model segmentation module 32 will first perform fine-grained deconstruction and semantic understanding of key components such as the neural network topology, knowledge meta-corpus, and reasoning execution flow. On this basis, the module will comprehensively analyze the dependencies, parameter redundancy, computational complexity, and other characteristics of different functional units within the model, and identify a series of functional substructures that can run independently and form a system. These substructures usually correspond to a core capability of the digital human, such as speech recognition, image classification, named entity recognition, sentiment analysis, etc. In the process of identifying substructures, the module will also fully absorb prior experience in the fields of natural language processing and knowledge engineering, actively explore the general construction paradigm of the digital human model, and strive to find the best balance between functional completeness and architectural simplicity. At the same time, the model segmentation module 32 will also actively connect with the actual application scenarios and functional requirements of local digital human services, widely listen to the insights of product managers and domain experts, and evaluate whether the boundary division of the functional submodel is reasonable from multiple perspectives. When a sub-model is not conceptually independent and complete enough, or has difficulty in exerting substantive value in application, the module will promptly adjust the segmentation strategy to allow a virtuous interaction between "technology-driven" and "demand-driven", and continuously iterate and optimize the connotation and extension of the functional sub-model with an open mind.

[0032] Based on the above embodiment, as an optional embodiment, the model segmentation module includes: a function module analysis unit and a sub-model segmentation unit; A functional module analysis unit is used to perform module function analysis on the local digital human model to obtain a functional topology diagram; Specifically, after obtaining the local digital human model, the functional module analysis unit will first conduct a static analysis on it. It will comprehensively examine the network architecture and weight distribution within the model, and use advanced algorithms such as graph neural networks and topic models to automatically summarize billions of neuron connections into dozens of functional clusters, and abstract and annotate the semantic features of each functional cluster. Generally speaking, the analysis unit will focus on the task-related layers within the model, such as the self-attention module responsible for semantic understanding in the encoder, the feedforward neural network responsible for knowledge assembly in the decoder, etc., and restore its role positioning in the overall cognitive reasoning process by analyzing its inherent computing logic and data flow. At the same time, the analysis unit will also focus on the domain-related parameters within the model, such as the adaptation layer weights obtained by fine-tuning for different vertical scenarios. By comparing the model changes before and after the analysis of the domain adaptation, the unit can identify the neural units that are strongly related to the specific scene, and then locate the key modules responsible for the knowledge storage and reasoning in the field. It is worth mentioning that in order to meet the needs of model analysis of different scales and complexities, the functional module analysis unit will also adaptively adjust the analysis algorithm. For models with fewer parameters and simpler architectures, the unit will adopt a fine-grained analysis strategy, by deeply examining the intermediate representation and attention distribution within the model, to reveal the complementarity and synergy of different modules in overall cognitive reasoning; while for complex models with large scale and many levels, the unit will adopt a coarse-grained analysis strategy, based on efficient algorithms such as graph embedding and spectral clustering to quickly lock key functional areas, so that computing resources will not be excessively consumed in the stripping of redundant modules. Through this analysis method that adapts to local conditions, the functional module analysis unit can efficiently and accurately restore the functional topology of the digital human model in various scenarios, laying a solid foundation for subsequent segmentation and integration.

[0033] The sub-model segmentation unit is used to segment the local digital human model according to the functional topology diagram to obtain multiple multi-granularity functional sub-models.

[0034] Specifically, after obtaining the functional topology diagram, the sub-model segmentation unit will immediately start the multi-granularity segmentation process. Relying on the module dependencies and interaction patterns presented in the diagram, the unit will automatically identify the key segmentation planes of the functional topology and implement recursive segmentation in both vertical and horizontal dimensions accordingly. In the vertical segmentation, the segmentation unit will be guided by the abstract hierarchy and adopt a top-down divide-and-conquer strategy to gradually simplify the overall digital human model into coarse-grained high-level modules. Generally speaking, these high-level modules include end-to-end core functions such as multi-round dialogue understanding, planning, and generation, which can support holistic services for specific scenarios and can be independently delivered to downstream industry users for direct call. On this basis, the segmentation unit will further implement horizontal segmentation. For highly cohesive and strongly coupled related modules, the unit will take interface consistency and reusability as the goal, adopt a bottom-up combine-and-conquer strategy, aggregate them into more compact coarse-grained sub-models, thereby achieving standardization in service encapsulation and interface specifications, and providing a basis for reuse and combination. For fine-grained modules with strong independence and high universal value, such as named entity recognition, relationship extraction, knowledge reasoning, etc. in specific vertical fields, the segmentation unit will encapsulate them atomically, so that each module focuses on specific functions, and flexibly supports various innovative combinations in the form of open interfaces, so that the service granularity and integration methods can achieve maximum flexibility according to industry needs. It is worth mentioning that in the segmentation process, the sub-model segmentation unit will also fully consider the differentiated characteristics of different sub-models and adaptively adjust the segmentation algorithm and evaluation indicators. For computationally intensive core modules, such as encoders and decoders, the segmentation unit will focus on their operating indicators such as latency and throughput, and use efficient pruning and distillation algorithms to slim down and ensure service performance; for knowledge-intensive vertical modules, such as vocabulary and corpora in specific fields, the segmentation unit will focus on evaluating their storage and retrieval efficiency, and minimize resource overhead through graph indexing, data structure optimization and other means. Through the segmentation strategy and evaluation mechanism that tailor-made for each student, the sub-model segmentation unit can ensure that each functional sub-model inherits key capabilities from the overall model while being lightweight and efficient, maintaining the same level of understanding, reasoning, and generation effects as the large model, and achieving the optimal balance between key indicators such as performance, power consumption, and generalization.

[0035] Function matching module 33, used to obtain the task requirements of the task to be deployed, and match the corresponding function sub-model according to the task requirements; Specifically, the function matching module 33 will continue to monitor task requests from local applications during daily operation. Once a new task is submitted, the module will immediately interact with the task initiator, and guide the other party to clarify its specific requirements through various methods such as graphical interface and natural language dialogue, including but not limited to: the field scenario, core task objectives, input data characteristics, response delay requirements, etc. After obtaining these structured task requirements, the module will immediately perform feature extraction and semantic analysis on them to extract the core requirements of the task at the cognitive interaction level. For example, when a user wants a digital human to assist him in browsing and analyzing a piece of text material, the module can automatically infer the urgent need for machine reading comprehension and opinion generation for this task. After clarifying the abstract features of the task, the function matching module 33 will use these features to quickly search and match similarities in the local digital human model library, striving to find those functional sub-models that can perfectly meet the task requirements. This matching process not only takes into account the task's requirements for model capabilities at the semantic level, but also fully estimates the task's consumption of computing resources at the spatiotemporal level, allowing the task and model to achieve a supply-demand balance in the two-dimensional space of "capability-resources". There will be no waste of resources due to excess model capabilities, nor will there be any task failure due to insufficient model capabilities.

[0036] Based on the above embodiment, as an optional embodiment, the function matching module includes: a task requirement decomposition unit and a function matching unit; Task requirement decomposition unit, used to decompose task requirements to obtain task functional requirements; Specifically, after receiving the task requirements submitted by the user, the task requirement decomposition unit will immediately start the multi-level semantic parsing process. The first is structural splitting. Faced with unstructured task descriptions of various forms, the unit will use natural language processing technology to convert them into standardized structured representations through methods such as syntactic analysis and chunk segmentation, and normalize key entities, attributes, and relationships based on the domain knowledge graph to ensure that subsequent functional inference can be carried out efficiently in a unified semantic space. On this basis, the decomposition unit will further carry out deep semantic analysis. Relying on the powerful understanding ability of the pre-trained language model, the unit can fully explore the implicit semantics behind the structured expression and capture the user's real demands in terms of task scenarios, constraints, and interaction methods. Generally speaking, the decomposition unit will focus on the following types of semantic information: the first is intention recognition, that is, analyzing the fundamental purpose that users want to achieve through tasks, such as information retrieval, question-and-answer dialogue, report generation, etc.; the second is topic recognition, that is, summarizing the core topics, key concepts, industry attributes, etc. involved in the task, revealing the user's field of interest; the third is constraint recognition, that is, refining the specific requirements of the task in terms of time, place, form, etc., so that the digital human service can be tailored. Through systematic semantic analysis, the decomposition unit can fully refine the explicit and implicit information in the unstructured task requirements, and summarize them into three categories of standardized semantic representations: intention, theme, and constraint, which constitute a comprehensive characterization of the functional requirements of digital human services. It is worth mentioning that in order to further improve the decomposition accuracy and efficiency, the task requirement decomposition unit also introduces a multi-dimensional reinforcement learning mechanism. On the one hand, the unit will continue to accumulate demand knowledge in different industries and fields through active learning, and incorporate it into the task decomposition model, so that the decomposition strategy can "hit the mark" with the help of domain common sense, and can also achieve fine-grained understanding in a few sample scenarios. On the other hand, the decomposition unit will continuously optimize the decomposition strategy itself through reinforcement learning, automatically learn the best practices of demand mining based on massive task training data, and achieve a dynamic balance between decomposition accuracy and generalization, so that digital humans can "learn once and provide inclusive services" and expand cognitive intelligence to long-tail needs. Through continuous learning and strategy evolution, the task requirement decomposition unit will build a new paradigm of demand-oriented functional representation, use artificial intelligence to open up a service closed loop from the demand side to the supply side, and make digital humans a new type of productivity that connects customers, industries, and industries, enabling high-quality economic development in a broader scenario.

[0037] The function matching unit is used to match the corresponding function sub-model according to the task function requirements.

[0038] Specifically, after obtaining the structured task function requirements, the function matching unit will immediately start a multi-stage matching process. The first is the initial screening based on semantic similarity. Faced with various functional requirements, the unit will first characterize the semantic features of the task function requirements and the candidate function sub-models based on the vector space model, and preliminarily judge the degree of matching between the two in terms of task type, subject area, interaction mode, etc. through classic algorithms such as cosine similarity and Euclidean distance, and quickly lock in potential target sub-models. On this basis, the matching unit will further carry out fine screening based on graph matching. Different from the independence assumption in the vector space model, there are often complex dependencies and associations between real tasks and function models. In order to accurately characterize this structured feature, the matching unit will map the semantic representation of the task function requirements and the candidate sub-models into a heterogeneous information network, and deeply explore the similarities between the two networks in topological structure and relationship semantics through graph matching algorithms such as node alignment and edge alignment, revealing the multi-dimensional implicit connection between task elements and model capabilities. Generally speaking, the matching unit will focus on the following types of structural features: the first is capability dependency, that is, the underlying functions and their combination relationships that need to be called for task implementation; the second is data dependency, that is, the input and output data types and their flow methods required for task execution; the third is constraint dependency, that is, the time, space, security, performance and other hard requirements of task definition. Based on the fine-grained representation of heterogeneous networks, the matching unit can conduct a full range of adaptability assessments on tasks and models in multiple dimensions such as semantics, structure, and constraints, to maximize the personalized and refined demands of industry needs. It is worth mentioning that in order to further improve the matching efficiency and effect, the functional matching unit also introduces multi-strategy reinforcement learning. On the one hand, the unit will continue to track the latest cutting-edge algorithms in the industry, such as graph neural networks and knowledge graph reasoning, and autonomously learn the pros and cons of various matching strategies through the strategy network, so that the matching process can adaptively select the optimal decision according to the characteristics of the task and the state of the environment, and achieve a dynamic balance between goals such as accuracy, efficiency, and generalization. On the other hand, the matching unit will continue to accumulate experience and knowledge in different industries and fields through active learning mechanisms. By incorporating feedback signals such as success stories and failure lessons into the strategy network, the matching model can examine the task requirements from the perspective of industry experts, integrate human value judgments and decision-making wisdom into the matching process, and truly make human-machine collaboration the "new normal" in the intelligent era. Through algorithm innovation and knowledge evolution, the functional matching unit has endowed the core capabilities of artificial intelligence such as perception, decision-making, and execution with distinct industry characteristics, allowing digital human services to fully meet the differentiated needs of different scenarios and become the key productivity of industrial intelligence in improving quality and efficiency.

[0039] The model reorganization module 34 is used to combine the corresponding functional sub-models into a target sub-model and deploy the target sub-model on the local device.

[0040] Specifically, after obtaining all the functional sub-models required for the target task, the model reorganization module 34 will immediately start a flexible orchestration process to build these sub-models into an organic whole. In this process, the module will first carefully sort out the input and output interfaces of each sub-model, extract their associated attributes in data format and semantic type, and build a dependency graph covering all sub-models. Based on this graph, the module can clearly depict the order and dependency conditions of each sub-model in the task execution process, and then generate an optimal task scheduling plan, so that each sub-model can achieve the closest coordination with the lowest coupling cost. After clarifying the scheduling plan, the module will also adopt differentiated integration strategies based on the heterogeneity of each sub-model. For homogeneous models with similar structures and common semantics, this module will implement deep fusion at the model level, and form a unified whole in the neural network topology through parameter sharing, joint training and other means; for heterogeneous models with different structures and domain-specific, this module will implement loose coupling at the interface level, and package them into standard functional units through microservices and containerization, give them a unified calling interface, and then let them perform their respective duties and operate in coordination by building a scheduling center and message bus. When all sub-models are integrated at the software level, the model reorganization module 34 will further perform a series of adaptability optimizations and enhancements on the integrated target sub-models, including parameter tuning, joint distillation, compilation acceleration, etc., in an effort to minimize the resource overhead of the target sub-model while ensuring performance, so that it can run smoothly in the local device.

[0041] The cloud-edge collaboration module 4 is used to establish a connection between the digital human model in the cloud and the target sub-model in the local device, and dynamically divide the task boundaries of the digital human model and the target sub-model according to the task characteristics of the tasks to be processed and the resource usage of the target sub-model, and assign tasks to the tasks to be processed according to the task boundaries.

[0042] Specifically, the cloud-edge collaboration module 4 first connects the target sub-model of the local device with the cloud-based digital human model through a high-speed and secure network channel to form an end-to-end collaborative reasoning link. On this basis, the module will perceive the status of cloud and local computing resources in real time, including parameters such as network latency, service load, and available computing power, and comprehensively analyze the type, data scale, computational complexity, and real-time response of the current tasks to be processed, and make a detailed portrait of the tasks. According to the task portrait and resource status, the cloud-edge collaboration module 4 will adaptively design an optimal task segmentation and offloading strategy, decompose the original task into multiple interrelated but relatively independent sub-tasks, and clearly divide the execution boundaries of each sub-task to form a task dependency and scheduling sequence diagram. For example, for complex tasks involving large-scale knowledge retrieval and deep semantic understanding, the cloud-edge collaboration module 4 will offload computationally intensive subtasks such as knowledge base query and long text analysis to the cloud and hand them over to the powerful digital human model for processing; for interactive tasks that are sensitive to latency, such as sentiment analysis, dialogue generation, and speech synthesis, the module will prioritize the execution of local target submodels, making full use of its millisecond-level reasoning capabilities to ensure real-time response. When a bottleneck occurs in local computing resources, the module will promptly offload overloaded tasks to the cloud to share and avoid a surge in response latency. In addition, the cloud-edge collaboration module 4 will continue to track the execution progress and service quality of each subtask, adjust the granularity and rhythm of task segmentation in real time according to the statistical characteristics of the task and the dynamic changes of resources, and seek the best balance between task granularity and scheduling overhead, local computing and load balancing, and strive to achieve the optimal coordination of end-to-end tasks with the lowest communication cost. It is worth mentioning that when allocating tasks, the cloud-edge collaboration module 4 will fully consider the privacy security and transmission efficiency of the data. For sensitive data involving user privacy, such as personal identification information, this module will ensure that it is only processed locally and will not be uploaded or leaked; for intermediate data that frequently interacts between the cloud and the edge, this module will adopt efficient data compression and encryption transmission mechanisms to minimize network overhead while ensuring communication security.

[0043] Based on the above embodiment, as an optional embodiment, the cloud-edge collaboration module 4 includes: a communication control module 41, a task scheduling module and a resource management module 42; The communication control module 41 is used to establish and manage the network connection between the cloud digital human model and the target sub-model in the local device, and control data transmission and synchronization; Specifically, the communication control module 41 will first comprehensively consider the deployment mode, network environment, security requirements and other factors of the digital human application scenario, and rationally plan the network topology and interconnection mode of the cloud and local. Generally speaking, for scenarios with high data security requirements and rich local computing resources, the module will give priority to deploying more models and computing tasks in local devices, and let the large cloud model only play a role of knowledge enhancement and computing power supplement, and only the necessary data link needs to be established between the two. For lightweight scenarios with strong universality and limited resources, the module tends to concentrate the main computing tasks and model capabilities in the cloud, and realize cognitive enhancement through real-time cloud-edge data intercommunication. When the cloud-edge topology is determined, the communication control module 41 will further select a suitable network transmission protocol to seek a balance between interactive real-time and communication efficiency. For example, for task scheduling and state synchronization with large data volumes, the module generally adopts a reliable transmission protocol based on TCP; while for delay-sensitive control instructions, lightweight protocols such as UDP will be selected.

[0044] The task scheduling module 42 is used to dynamically divide the task boundaries of the digital human model and the target sub-model according to the task characteristics of the task to be processed and the preset resource constraints, and to allocate the task to be processed according to the task boundaries; Specifically, after receiving the pending task request from the local digital human module, the task scheduling module 42 will first conduct a comprehensive analysis of the inherent characteristics of the task. The task characteristics here mainly include but are not limited to: the computational complexity of the task, the required knowledge domain, the degree of dependence on modal fusion capabilities, the sensitivity to local information, the relevance to the user's long-term portrait, the degree of demandingness of the real-time interactive experience, etc. After qualitatively and quantitatively characterizing these characteristics, the module will intelligently profile the pending tasks based on this, and evaluate the degree of demandingness of key performance indicators such as local computing resources, network bandwidth, privacy security, and interaction delay, which will serve as an important basis for subsequent task allocation. At the same time, the task scheduling module 42 will also monitor the resource utilization of the cloud and local in real time, including the computing power level of the cloud GPU cluster, the load intensity of the local chip, the network delay between the cloud and the edge, the remaining capacity of the local storage, etc., which will serve as the red line and benchmark for dividing the task boundary. In addition, the module also allows the application or user to set global resource constraints, such as requiring the overall response time to be no more than 1 second, or the proportion of tasks taken over by the cloud to be no more than 30%. These constraints will further limit the flexibility of task scheduling.

[0045] The resource management module 43 is used to monitor the resource usage of the cloud and the local resources.

[0046] Specifically, the resource management module 43 will manage all kinds of software and hardware resources in the cloud data center and local smart terminals throughout their life cycle. The resources here mainly include but are not limited to: the computing power scale and utilization rate of the cloud GPU / TPU cluster, the computing power level and power consumption budget of the local chip, the storage capacity and retrieval efficiency of the cloud knowledge base, the remaining space and data throughput of the local memory / flash memory, the delay jitter and packet loss rate of the network link between the cloud and the edge, etc. For these resource data, the module will actively collect them regularly through remote calls to form a panoramic view of the resources, and will also passively receive resource alarm events reported by the cloud edge to promptly discover potential resource bottlenecks. It is worth mentioning that the resource management module 43 not only pays attention to the static capacity and configuration of resources, but also pays more attention to their dynamic usage intensity and load changes. For core computing resources such as GPUs, this module will closely monitor the dynamic changes of their computing power levels as tasks are executed, and evaluate whether the load is saturated and whether there are fragmentation problems; for storage resources that are easily restricted, such as memory, this module will count the idle rate in real time to determine whether there are signs of performance degradation such as frequent data exchange and reduced throughput; for communication resources that are prone to congestion, such as network bandwidth, this module will continuously monitor QoS indicators such as network latency and packet loss rate, and evaluate the link quality and stability between the cloud and edge. In addition, to support the resource constraint solution of the scheduling module 42, this module can also perform predictive analysis on the medium- and long-term use of cloud-edge resources, fit the changing trend of resource use through time series, regression and other algorithms, identify possible resource bottlenecks in advance, and calmly respond to dynamic adjustments to the scheduling plan.

[0047] Based on the above embodiment, as an optional embodiment, the system further includes: a data security module; The data security module is used to encrypt the data transmitted between the cloud digital human module and the local digital human module.

[0048] Specifically, the data security module plays a key role in all stages of cloud-edge collaboration. The first is network transmission encryption. In the face of sensitive data transmitted in the public network, the module will use industry standard encryption algorithms such as RSA and AES, and combine with the key regular update mechanism to perform end-to-end encryption on the transmission content, ensuring that the data is difficult to crack even if it is intercepted, and preventing risks such as network eavesdropping and man-in-the-middle attacks from a physical level. At the same time, the module will also use technical means such as digital signatures and timestamps to embed verification information in the encrypted data, so that once the data is tampered with, it can be quickly detected, further improving the integrity and non-repudiation of the transmission process. In the data landing stage, the security module will also perform secondary encryption on the statically stored privacy data, using new cryptographic technologies such as attribute-based encryption (ABE) to strictly limit the access rights of different subjects, so that sensitive information can be "traceable" and prevent illegal access by internal personnel. Generally speaking, the encryption granularity can be refined to the field level, and the access policy can express complex logical combinations, such as "the department belongs to the technology center and the rank is not lower than the director", so that data sharing can be scaled in a trustworthy and controllable manner. It is worth mentioning that in order to carry out data analysis under high-intensity encryption, the module also supports a variety of privacy computing technologies. Through innovative solutions such as homomorphic encryption and secure multi-party computing, the data security module can directly perform statistical analysis on data in the ciphertext space, and mine the value of data without decrypting or leaking individual privacy, allowing data elements to circulate in protection and create value in circulation.

[0049] Based on the above embodiment, as an optional embodiment, the system further includes a monitoring log module; The monitoring log module is used to monitor and analyze the performance indicators, error logs and access logs of the cloud digital human module and the local digital human module in real time.

[0050] Specifically, the monitoring log module adopts a full-stack, three-dimensional monitoring method. The first is basic layer monitoring. Faced with massive, heterogeneous software and hardware resources under the cloud-edge-end collaborative architecture, the module makes full use of virtualization, containerization and other technical means to collect IaaS-level operating indicators such as CPU, memory, disk, and network with extremely fine granularity and extremely fast frequency, and intelligently analyzes potential risks through threshold alarms, anomaly detection and other algorithms, allowing the operation and maintenance team to perceive and deal with potential faults with seconds of efficiency. On this basis, the module also provides service monitoring at the PaaS layer. Unlike IaaS, which focuses on a single resource, this module examines the service status of the system from an end-to-end perspective. Through distributed tracking, call chain analysis and other means, the monitoring module can accurately locate the bottleneck resources and fault points behind common problems such as service avalanches and interface timeouts, and then achieve rapid recovery and long-term optimization of problems through technical means such as service orchestration and self-healing to ensure business continuity. It is worth mentioning that at the SaaS layer, the monitoring module also supports all-round business monitoring. By collecting business data such as user complaints and combining it with intelligent algorithms such as semantic analysis and sentiment analysis, the monitoring module can quickly and comprehensively gain insight into business shortcomings and room for improvement, allowing digital people to enhance their development potential in improving user experience. At the same time, through data mining, causal analysis and other technologies, the monitoring module can summarize the key performance indicators that best reflect the core demands of the enterprise from thousands of business indicators, allowing system optimization and business development to form a positive closed loop.

[0051] Based on the above embodiment, as an optional embodiment, the system further includes: a model management module; The model management module is used to manage the registration, query, update, and offline of the cloud digital human module and the local digital human module.

[0052] Specifically, the model management module adopts a closed-loop management mode for the entire life cycle. The first is the model registration link. Faced with a large number of models scattered among various development teams, this module provides a unified model registration entry. R&D personnel only need to submit the model's source code, usage documents and other key information according to the specifications, and the platform will automatically assign it a unique identifier for subsequent management and use. At the same time, the registration process is also a "physical examination". The system automatically scans and analyzes indicators such as code quality and resource consumption to timely discover and feedback potential performance risks, so that the online model can reach the best state. On the basis of model registration, the management module also supports flexible model query functions. Through labeling and semantic retrieval methods, users can quickly locate model resources in specific fields and scenarios, match business needs in a targeted manner, and greatly improve resource reuse efficiency. At the same time, the management module also has built-in multiple performance evaluation frameworks. Through standardized test cases, the system can objectively evaluate the performance level of the model under different hardware and different working conditions, so that the right model can be used in the right place. It is worth mentioning that in order to meet the actual needs of algorithm iteration, the management module also provides a flexible model update mechanism. On the one hand, developers can easily submit new versions of the model. The system controls iteration risks through online A / B testing, canary release and other mechanisms, allowing innovative results to quickly serve users. On the other hand, old versions are also archived in the form of container images, which can be rolled back at any time with one click, so that model releases can take into account both agility and stability. Finally, for models that are no longer in use, the system also provides a standardized offline process. Through built-in governance rules, the management module regularly evaluates the usage status of the model, automatically archives long-term idle "sedimentation" resources, and saves computing power costs to the maximum extent. At the same time, for vertical industry models involving sensitive data, the system also embeds a secure erase function to destroy model parameters by cryptographic means to maximize data security.

[0053] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure.

[0054] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A lightweight digital human system, characterized in that: The system includes: a cloud digital human module, a local digital human module, a local digital human model deployment module and a cloud-edge collaboration module; The cloud digital human module is used to deploy the trained digital human model in the cloud; The local digital human module is used to compress and optimize the digital human large model to obtain a target local digital human model, and to functionally split and reorganize the target local digital human model according to local device requirements to obtain a target sub-model; The local digital human model deployment module is used to deploy the target sub-model on a local device; The cloud-edge collaboration module is used to establish a connection between the digital human model in the cloud and the target sub-model in the local device, and dynamically divide the task boundaries of the digital human model and the target sub-model according to the task characteristics of the task to be processed and the resource usage of the target sub-model, and assign tasks to the tasks to be processed according to the task boundaries.

2. The system according to claim 1, characterized in that The cloud-based digital human module includes: a digital human large model training module and a cloud-based deployment module; A digital human model training module is used to construct a preliminary digital human model and a sample training set, and train the preliminary digital human model based on the sample training set. When the training result reaches a preset standard, the trained digital human model is obtained; The cloud deployment module is used to encapsulate the trained digital human model into a standardized cloud service.

3. The system according to claim 1, characterized in that The local digital human module includes: a model compression module, a model segmentation module, a function matching module and a model reorganization module; The model compression module is used to obtain a digital human model to be deployed, and compress and optimize the digital human model to obtain a local digital human model; The model segmentation module is used to segment the local digital human model to obtain multiple multi-granularity functional sub-models; The function matching module is used to obtain the task requirements of the task to be deployed, and match the corresponding function sub-model according to the task requirements; The model reorganization module is used to combine the corresponding functional sub-models into a target sub-model and deploy the target sub-model on a local device.

4. The system according to claim 1, characterized in that The cloud-edge collaboration module includes: a communication control module, a task scheduling module and a resource management module; The communication control module is used to establish and manage the network connection between the digital human model in the cloud and the target sub-model in the local device, and control data transmission and synchronization; The task scheduling module is used to dynamically divide the task boundaries of the digital human model and the target sub-model according to the task characteristics of the task to be processed and the preset resource constraints, and to allocate the task to be processed according to the task boundaries; The resource management module is used to monitor the resource usage of the cloud and the local resources.

5. The system according to claim 3, characterized in that The model compression module includes: a knowledge distillation unit, a model quantization unit, a model pruning unit and a local optimization unit; the compression and optimization of the digital human model to obtain a local digital human model includes: The knowledge distillation unit is used to construct a first local digital human model, and use the digital human large model as a teacher model to train the first local digital human model to obtain a second local digital human model; The model quantization unit is used to quantize the weight and activation of the second local digital human model to obtain a third local digital human model; The model pruning unit is used to prune the third local digital human model to obtain a fourth local digital human model; The local optimization unit is used to compress and convert the fourth local digital human model into a format supported by the mobile terminal inference engine to obtain the local digital human model.

6. The system according to claim 3, characterized in that The model segmentation module includes: a functional module analysis unit and a sub-model segmentation unit; The functional module analysis unit is used to perform module function analysis on the local digital human model to obtain a functional topology diagram; The sub-model segmentation unit is used to segment the local digital human model according to the functional topology diagram to obtain the multiple multi-granularity functional sub-models.

7. The system according to claim 3, characterized in that The function matching module includes: a task requirement decomposition unit and a function matching unit; The task requirement decomposition unit is used to decompose the task requirement to obtain task function requirements; The function matching unit is used to match the corresponding function sub-model according to the task function requirements.

8. The system according to claim 1, characterized in that The system further comprises: a data security module; The data security module is used to encrypt the data transmitted between the cloud digital human module and the local digital human module.

9. The system according to claim 1, characterized in that The system also includes a monitoring log module; The monitoring log module is used to monitor and analyze the performance indicators, error logs and access logs of the cloud digital human module and the local digital human module in real time.

10. The system according to claim 1, characterized in that The system also includes: a model management module; The model management module is used to manage the registration, query, update and offline of the cloud digital human module and the local digital human module.

Images